Restoring Mycenaean Linear B ‘A&B’ series tablets using supervised and transfer learning

Katerina Papavassileiou, Dimitrios KosmopoulosACL 2024

TL;DR

This paper explores using machine learning, specifically Bidirectional Recurrent Neural Networks (BRNNs) and transfer learning, to restore damaged Mycenaean Linear B clay tablets. By training models on existing tablet data and then fine-tuning them, the researchers aim to improve the accuracy of predicting missing characters in these ancient texts, addressing the challenge of limited data.

What Problem Does This Paper Address?

Mycenaean Linear B, a syllabic script used in ancient Greece (c. 1400-1200 B.C.), presents a significant challenge for restoration due to the scarcity of data and the poor preservation state of many tablets (broken, worn, or burnt). While previous work has attempted to restore these texts, often relying on expert knowledge or simpler statistical models, this paper aims to leverage advanced machine learning techniques, including transfer learning, to improve the accuracy and efficiency of this process. The administrative nature of the tablets and the presence of both phonetic (syllabograms) and non-phonetic (ideograms) symbols add further complexity.

Methods

The core of the methodology involves using a Bidirectional Recurrent Neural Network (BRNN) for character-level text infilling. The researchers experimented with two main approaches:

  • Supervised Learning: Training BRNN models from scratch on datasets derived from Mycenaean tablets. A key variation involved training models with and without ideograms to assess their impact.
  • Transfer Learning (TL): This was a primary focus. Two TL techniques were explored:
  • Weight Initialization (CASE A): Using the optimal parameters from a BRNN model trained on a larger dataset (series D) as the initial weights for training a model on a smaller dataset (series A&B).
  • Freezing Pre-trained Layers (CASE B): Freezing the weights of the pre-trained model (from series D) and training an additional hidden layer for the target dataset (series A&B).

The evaluation used a Leave-One-Out Cross-Validation (LOOCV) approach to account for data scarcity.

Datasets

The study utilized two main datasets derived from Mycenaean Linear B tablets, both originating from Knossos:

  • Series D: This is the larger dataset, comprising around 1100 tablets related to sheep herd accounts. For experiments, 513 complete sequences were extracted, augmented to 2565 for training. A modified version of this dataset was created by removing ideograms, numeric signs, and measurement units.
  • Series A&B: This is a smaller dataset, consisting of tablets related to staff lists and personnel situations. 426 complete sequences were gathered, with 159 sequences from damaged tablets. This dataset was used as the target for transfer learning.

Key Insights and Ideas

  • Ideograms' Impact: Removing ideograms from the sequences led to a small improvement in BRNN model performance for Series D, suggesting that focusing solely on phonetic components can be beneficial.
  • Transfer Learning Potential: While TL did not consistently outperform training from scratch on synthetic gaps, it showed promise in handling real, complex cases where supervised learning struggled. This suggests TL can offer complementary solutions.
  • Data Scarcity is Key: The significant difference in performance between Series D and Series A&B highlights the critical impact of dataset size and sequence length on model effectiveness.
  • Expert Collaboration: The study validates the model's predictions against expert annotations, finding that TL models sometimes align better with expert opinions in challenging cases.

Key Results

  • Series D (without ideograms): Achieved slightly better scores (e.g., TOP-20 accuracy of 80.12%) compared to training with ideograms (78.17%).
  • Series A&B (supervised): Performance was lower than Series D, with a TOP-20 accuracy of 66.20%, attributed to fewer and shorter sequences.
  • Transfer Learning (CASE A & B):
    • For synthetic gaps, TL did not show overall improvement compared to training from scratch.
  • For real cases, TL models (especially CASE B) demonstrated better performance in some complex scenarios, successfully providing solutions where supervised models failed. For instance, in 29 real cases, TL models achieved higher agreement with expert predictions in TOP-5 and TOP-10 rankings compared to the baseline BRNN.

Limitations and Future Work

  • Data Size: The primary limitation acknowledged is the scarcity of data, particularly for Series A&B, which impedes robust training and augmentation.
  • Sequence Length: Shorter sequences in Series A&B offer less context, making prediction more challenging.
  • Visual Modality: The current models do not incorporate visual information from the tablets, which experts use extensively. The authors suggest this as a crucial area for future research.
  • More Series and Sites: Expanding the dataset to include more series and tablets from other sites (Pylos, Thebes, etc.) is proposed to increase robustness and diversity.
  • Further TL Investigation: Continued research into TL methods is recommended, especially with the potential for increased data.

Summary updated

BibTeX
@inproceedings{papavassileiou2024restoring,
  author = {Katerina Papavassileiou and Dimitrios Kosmopoulos},
  title = {{Restoring Mycenaean Linear B ‘A\&B’ series tablets using supervised and transfer learning}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2024},
}
Figure